How AGI Will Change Business

The interesting question is not what AGI will do to business. Nobody can answer that, and the people claiming to have a date disagree with each other by twenty years. The answerable question is what substantially more capable AI is doing to business right now, because that record already contains the pattern.

The number nobody wants to look at

McKinsey's global survey, 1,719 respondents across 97 nations, fielded May to June 2026: enterprise-scale adoption rose from 38% to 44%. The share of organisations attributing any EBIT impact to AI stayed at 37%, unchanged year on year. The share qualifying as high performers, meaning at least 5% of EBIT attributable to AI, stayed at 6%. Also flat.

Meanwhile 80% of individuals report improved productivity. Individual gains are real and they are not aggregating into company results.

The central-bank evidence is blunter. A survey of roughly 6,000 CEOs, CFOs and senior finance managers across the US, UK, Germany and Australia, run by the Atlanta Fed, the Bank of England and the Bundesbank, found nine in ten executives reporting no impact on employment or productivity over the past three years, with realised productivity gains totalling 0.29%.

Executives in a separate survey claimed 1.8% productivity growth from AI in 2025. The gains implied by their own reported revenue and employment figures were 0.6%.

Be careful which adoption number you quote

A Federal Reserve Board note from April 2026 reconciled three US surveys that appear to contradict each other. Firm-weighted adoption: 18%. Individual work-related use: 41%. Employment-weighted firm adoption: 78%.

None of them is wrong. They count different things. Anyone quoting "X% of businesses use AI" without saying how it was weighted is producing a number that cannot be acted on.

The Census Bureau's firm-level work adds the detail that matters more than the headline: 57% of adopting firms use AI in three or fewer business functions, and 65% of those using it at task level restrict it to three or fewer tasks. Adoption is a mile wide and an inch deep.

What separates the 6% from everyone else

This is where the evidence is unusually clear, and unusually actionable. McKinsey found 73% of high performers are fundamentally redesigning workflows, against 25% of everyone else. Not buying more tools. Changing how the work runs.

A field experiment on 515 high-growth startups tested this directly. The researchers call the barrier the mapping problem: firms cannot work out where in their own production process AI creates value. The treatment was not technology. Treated firms were simply given information about how other firms had reorganised production around AI. Results against control:

• 44% more AI use cases discovered
• 12% more tasks completed
• 18% more likely to acquire paying customers
• 1.9x higher revenue
• Capital investment demand down 39.5%, labour demand unchanged

An information-only intervention produced nearly double the revenue. The binding constraint is organisational knowledge, not model capability.

Where value is actually captured

Goldman Sachs analysed Q4 earnings calls and found no meaningful economy-wide relationship between productivity and AI adoption. Seventy per cent of S&P 500 firms discussed AI; only 10% quantified its impact on any specific use case, and only 1% quantified its impact on earnings.

Two use cases showed a median ~30% productivity gain: customer support and software development. Those are precisely the two with the strongest academic evidence behind them. In a study of over 5,000 customer support agents, generative AI produced 14% more issues resolved per hour on average, 34% for novices and close to nothing for the most experienced.

That heterogeneity is the recurring finding across the literature. AI raises the floor far more than it raises the ceiling.

The cost curve is not the story people think

Inference cost per token collapsed: GPT-3.5-equivalent performance fell from $20 per million tokens in November 2022 to $0.07 by October 2024, more than a 280-fold reduction.

Cost per completed task has not fallen anything like as far, because reasoning and agentic workflows consume vastly more tokens per task. Epoch AI estimates token consumption growing around 10x per year against inference capacity growing 3.4x, and expects that to push frontier prices up. McKinsey found 20% of organisations saying AI operating costs have actively limited their adoption. A cost collapse that never reaches the P&L is exactly what you would expect when consumption expands to absorb it.

How much of this is real, and how fast

Serious economists disagree by roughly a factor of twenty. Daron Acemoglu estimates total factor productivity gains of under 0.53% over ten years. The OECD estimates 0.4 to 1.3 percentage points per year over ten years for high-exposure G7 economies. The divergence comes from assumptions about task exposure and adoption pace, not from disputed data.

The most honest single figure available: generative AI now saves US workers 2.2% of work hours, up from 1.6% in late 2024. Real, measurable, compounding, and nowhere near the rhetoric.

What a smaller company should take from this

The adoption gap by size is large. Firms with 250+ employees: 37%. Firms under 20 people: below 20%, with no significant change over six months of measurement. Large organisations scaling agents: 40% and rising. Smaller ones: 22% and flat.

But the thing that separates winners from everyone else is workflow redesign, and that is the one advantage a small team actually has. A twelve-person studio can restructure how a job runs in a week. A large organisation takes a year and a change-management programme. The evidence says the constraint is knowing where to apply it, not affording it.

So the practical move is not a bigger AI budget. It is picking one workflow you run repeatedly, mapping where the time actually goes, redesigning that process around what these tools reliably do, and measuring cost per completed job before and after. That is what the 6% did.